v0.0.24

Important

Superseded by v0.0.25. The v0.0.24 packaging run failed and this release has no binary asset. Use v0.0.25 instead.

Highlights

• Adds the Nemotron VoiceChat 11B full-duplex MLX runtime and Soniqo terminal voice-chat demo, with model-native turn taking, configurable MCP tools, Apple Reminders integration, and structured latency and acoustic-quality benchmarks.
• Adds MOSS, Cohere, Voxtral, Canary, and native Nemotron speech-recognition paths, plus timestamped audio capture and LocalVQE acoustic echo cancellation.
• Adds the CSM speech runtime, an MLX DeepFilterNet3 engine, SpeechBrain ECAPA language identification, and substantial Qwen and Gemma inference performance and reliability improvements.

What's Changed

• docs: list AnythingLLM compatibility by @ivan-digital in #417
• Stride-aware Sortformer output reads and an experimental Ultra 8-speaker variant by @ivan-digital in #418
• Add timestamped independent audio capture by @ivan-digital in #419
• Add match-only ReDimNet embeddings for short utterances by @ivan-digital in #421
• feat: add LocalVQE acoustic echo cancellation by @ivan-digital in #422
• fix: prevent LocalVQE CoreML buffer exhaustion by @ivan-digital in #424
• Add native MOSS Transcribe CoreML runtime by @ivan-digital in #426
• Add Cohere Transcribe and Voxtral Mini ASR by @ivan-digital in #423
• Report diarization DER by speaker count by @ivan-digital in #420
• Add CSM (Conversational Speech Model) runtime by @ivan-digital in #425
• Add long-context MLX runtime for MOSS by @ivan-digital in #427
• Add native Nemotron MLX streaming runtime by @ivan-digital in #428
• feat: add MLX engine for DeepFilterNet3 speech enhancement by @ivan-digital in #430
• Fix microphone resampling drift by @ivan-digital in #429
• Fix Qwen3-TTS model downloads by @mchernyavskaya in #431
• Consolidate model downloads onto one transfer engine by @ivan-digital in #433
• Read chat model quantization from the checkpoint instead of assuming INT4 by @ivan-digital in #434
• Default the Qwen3.5 chat MLX loader to INT5 instead of INT4 by @ivan-digital in #435
• Band Gemma 4's sliding attention instead of masking a full score matrix by @ivan-digital in #436
• Order adopters by stars across every README by @ivan-digital in #437
• Merge Gemma 4's BPE through a queue instead of rescanning every pair by @ivan-digital in #438
• Sample Gemma 4 on the GPU instead of copying the vocabulary per token by @ivan-digital in #439
• Append Gemma 4's KV cache in place and bound its sliding layers by @ivan-digital in #440
• Set the sliding-window logit bound from what blocking actually costs by @ivan-digital in #441
• Add Canary ASR, and hold the mel front end to the training contract by @ivan-digital in #446
• Make Qwen async generation cooperatively cancellable by @ThorFuchs in #444
• Add deterministic local Qwen3-TTS loading by @ThorFuchs in #442
• Route Qwen diagnostics through unified logging by @ThorFuchs in #443
• Add complete VoiceChat 11B duplex speech-to-speech runtime for MLX by @ivan-digital in #447
• Make VoiceChat INT5 faster than real time by @ivan-digital in #449
• Compile Qwen3-TTS code predictor frames by @ivan-digital in #450
• Bound the MLX buffer pool in Qwen3TTS E2E suites by @enrzh in #448
• Add SpeechBrain ECAPA language identification by @ivan-digital in #451
• Give VoiceChat real-time inference headroom by @ivan-digital in #452
• Add full-duplex VoiceChat CLI with native MCP tools by @ivan-digital in #455

New Contributors

• @mchernyavskaya made their first contribution in #431
• @ThorFuchs made their first contribution in #444
• @enrzh made their first contribution in #448

Full Changelog: v0.0.23...v0.0.24